When Platforms Say ‘Don’t Optimize,’ Good Groups Run Experiments

A fast be aware up entrance, so we begin on the proper foot.

The analysis I’m about to reference is not mine. I did not run these experiments. I’m not affiliated with the authors. I’m not right here to “endorse” a camp, choose a facet, or crown a winner. What I’m going to endorse, loudly and with out apology, is measurement. Replication. Actual-world experiments. The form of work that teaches us in actual time, in actual life, what modifications when an LLM sits between prospects and content material. We’d like extra examined information, and this is a kind of beginning factors.

In case you do nothing else with this article, do that: Learn the paper, then run your personal check. Whether or not your outcomes agree or disagree, publish them. We’d like extra receipts and fewer scorching takes.

Now, the motive I’m scripting this.

Over the final yr, the trade has been pushed towards a neat, comforting story: GEO is simply search engine optimization. Nothing new to study. No want to change how you’re employed. Simply maintain doing the fundamentals, and all the pieces shall be positive.

I don’t purchase that.

Not as a result of SEO fundamentals stopped mattering. They nonetheless matter, they usually stay mandatory. However as a result of “mandatory” is not the identical as “adequate,” and since the incentives behind platform messaging do not at all times align with the operational realities companies are strolling into and coping with.

The Narrative And The Incentives

In case you’ve paid consideration to public steering coming from the main search platforms these days, you’ve most likely heard a model of: Don’t focus on chunking. Don’t create “bite-sized chunks.” Don’t optimize for the way the machine works. Focus on good content material.

That’s been echoed and amplified throughout trade protection, although I would like to be exact about my place right here. I’m not claiming a conspiracy, and I’m not saying anybody is being deliberately deceptive. I’m not doing that.

I’m saying one thing a lot easier. It’s my opinion and occurs to be based mostly on precise expertise – when messaging repeats throughout a number of spokespeople in a decent window, it alerts an inner alignment effort.

That’s not an insult nor is it an ethical judgment. That’s how massive organizations function when they need the market to hear one clear message. I used to be a part of precisely that sort of surroundings for properly over a decade in my profession.

And the message itself, on its face, is not incorrect. You’ll be able to completely harm your self by over-optimizing for the incorrect proxy. You’ll be able to completely create brittle content material by attempting to sport a system you do not absolutely perceive. In lots of circumstances, “write clearly for people” is strong baseline steering.

The issue is what occurs when that baseline steering turns into a blanket dismissal of how the machine layer works in the present day, even when it’s unintentional. As a result of we are not in a “10 blue hyperlinks” world anymore.

We are in a world the place reply surfaces are increasing, search journeys are compressing, and the unit of competitors is shifting from “the web page” to “the chosen portion of the web page,” assembled into a solution the consumer by no means clicks previous.

And that is the place “GEO is simply search engine optimization” begins to break in my thoughts.

The Flawed Query: “Is Google Nonetheless The Greatest Visitors Driver?”

Executives love comforting statements: “Google nonetheless dominates search. Conventional search engine optimization nonetheless drives the most visitors. Subsequently this LLM-stuff is overblown.”

The first half is true, however the conclusion is the place corporations get harm.

The largest danger right here is asking the incorrect query. “The place does visitors come from in the present day?” is a dashboard query, and it’s backward-looking. It tells you what has been true.

The extra necessary questions are forward-looking:

• What occurs to your small business when discovery shifts from clicks to solutions?
• What occurs when the buyer’s journey ends on the outcomes web page, inside an AI Overview, inside an AI Mode expertise, or inside an assistant interface?
• What occurs when the platform retains the consumer, monetizes the reply floor, and your content material turns into a supply enter quite than a vacation spot?

In order for you the conduct trendline in plain phrases, begin right here, with the 2024 SparkToro study, then check out what Danny Goodwin wrote in 2024, and as a follow-up in 2025 (spoiler – zero click on cases elevated Y-o-Y). And whereas some sources are a few years outdated, you may simply discover newer information displaying the development rising.

I’m not utilizing these sources to declare “the sky is falling.” I’m utilizing them to reinforce a easy operational actuality: If the click on declines, “rating” is now not the finish purpose. Being chosen into the reply turns into the finish purpose.

That requires extra considering past traditional search engine optimization. Not as an alternative of it. On high of it.

The Platform Footprint Is Altering, And The Enterprise Mannequin Is Following

In order for you to perceive why the public messaging is conservative, you will have to have a look at the platform’s strategic route.

Google, for instance, has been increasing AI reply surfaces, and it’s not delicate. Each AI Overviews and AI Mode noticed bulletins of enormous expansions throughout 2025.

Once more, discover what this means at the working stage. When AI Overviews and AI Mode broaden, you’re not simply coping with “rating alerts.” You’re coping with an expertise layer that may reply, summarize, suggest, and route a consumer and not using a click on.

Then comes the half everybody pretends not to see till it’s unavoidable: Monetization follows consideration.

This is now not hypothetical. Search Engine Journal lined Google’s official rollout of ads in AI Overviews, which issues as a result of it alerts this reply layer is being handled as a sturdy interface floor, not a short lived experiment.

Google’s personal Adverts documentation reinforces the identical level: This isn’t simply “one thing individuals observed,” it’s a supported placement sample with actual operational steering behind it. And Google noted mid-last-year that AI Overviews monetize at an identical price to conventional search, which is a quiet sign that this isn’t a facet function.

You do not want to be cynical to learn this clearly. If the reply floor turns into the major floor, the advert floor will evolve there too. That’s not a scandal a lot as simply the actuality of the place the mannequin is evolving to.

Now join the dots again to “don’t focus on chunking”-style steering.

A platform that is actively increasing reply surfaces has a number of official causes to discourage the market from “engineering for the reply layer,” together with high quality management, spam prevention, and ecosystem stability.

Companies, nevertheless, do not have the luxurious of optimizing for ecosystem stability. Companies should optimize for enterprise outcomes. Their very own outcomes.

That’s the stress.

This isn’t about blaming anybody. It’s about understanding misaligned targets, so that you don’t make selections that really feel secure however value you later.

Discovery Is Fragmenting Past Google, And Early Alerts Matter

I’m on report that conventional search is nonetheless an necessary driver, and that optimizing on this new world is additive, not an in a single day alternative story. However “additive” nonetheless modifications the workflow.

AI assistants are turning into measurable referrers. Not dominant, not decisive on their very own, however significant sufficient to observe as an early indicator. Two examples that seize this development.

TechCrunch famous that whereas it’s not sufficient to offset the lack of visitors from search declines, information websites are seeing progress in ChatGPT referrals. And Digiday has information displaying visitors from ChatGPT doubled from 2024 to 2025.

Why do I embrace these?

As a result of this is how platform shifts look in the early phases. They begin small, then they change into regular, then they change into default. In case you look forward to the “huge numbers,” you’re late constructing competence and in taking motion. (Keep in mind “directories”? Yeah, Search ate their lunch.)

And competence, on this new surroundings, is not “how do I rank a web page.” It’s “how do I get chosen, cited, and trusted when the interface is an LLM.”

This is the place the “GEO is simply search engine optimization” framing stops being a useful simplification and begins turning into operationally harmful.

Now, The Receipts: A Paper That Assessments GEO Techniques And Reveals Measurable Variations

Let’s discuss the analysis. The paper I’m referencing right here is publicly available, and I’m going to summarize it in plain English, as a result of most practitioners do not have time to parse educational construction throughout the week.

At a excessive stage, the (“E-GEO: A Testbed for Generative Engine Optimization in E-Commerce”) paper assessments whether or not frequent human-written rewrite heuristics truly enhance efficiency in an LLM-mediated product choice surroundings, then compares that to a extra systematic optimization strategy. It makes use of ecommerce as the proving floor, which is good for one motive: Outcomes may be measured in ways in which map to cash. Product rank and choice are economically significant.

This is necessary as a result of the GEO dialog typically will get caught in “vibes.” In distinction, this work is attempting to quantify outcomes.

Right here’s the key punchline, simplified:

A whole lot of frequent “rewrite recommendation” does not assist on this surroundings. A few of it may be impartial. A few of it may be destructive. However after they apply a meta-optimization course of, prompts enhance constantly, and the optimized patterns converge on repeatable options.

That convergence is the half that ought to make each practitioner sit up. As a result of convergence suggests there are secure alerts the system responds to. Not mystical. Not magical. Not purely random.

Secure alerts.

And this is the place I come again to my earlier level: If GEO have been really “simply search engine optimization,” then you definately would anticipate traditional human rewrite heuristics to translate cleanly. You’d anticipate the successful playbook to be acquainted.

This paper suggests the actuality is messier. Not as a result of search engine optimization stopped mattering, however as a result of the unit of success modified.

• From web page rating to reply choice.
• From persuasion copy to choice copy.
• From “learn the complete web page” to “retrieve the finest phase.”
• From “the consumer clicks” to “the machine chooses.”

What The Optimizer Retains Discovering, And Why That Issues

I would like to watch out right here, as I’m not telling you to deal with this paper like doctrine. It’s best to not settle for it on face worth and immediately undertake this as gospel. It’s best to deal with it as a public experiment that deserves replication.

Now, the most dear output isn’t the precise numbers of their surroundings, however quite, it’s the form of the resolution the optimizer retains converging on. (The title of their system/course of is optimizer.)

The optimized patterns repeatedly emphasize readability, explicitness, and decision-support construction. They scale back ambiguity. They floor constraints. They outline what the product is and is not. They make comparisons simpler. They encode “selection-ready” information in a type that is simpler for retrieval and rating layers to use.

That is a distinct purpose than traditional advertising and marketing copy, which frequently leans on narrative, model really feel, and emotional persuasion.

These issues nonetheless have a spot. However if you’d like to be chosen by an LLM appearing as an middleman, the content material wants to do a second job: change into machine-usable choice assist.

That’s not “anti-human.” It’s pro-clarity, and it’s the form of element that may come to outline what “good content material” means in the future, I believe.

The Common LLM-Optimization Rewrite Recipe, Framed As A Reusable Template

What follows is not me inventing a course of out of skinny air. This is me reverse-engineering what their optimization course of converged towards, and turning it right into a repeatable template you may apply to product descriptions and different decision-heavy content material.

Deal with it as a place to begin, then check it. Revise it, create your personal model, no matter.

Step 1: State the product’s function in a single sentence, with express context.
Not “premium high quality.” Not “finest in school.” Goal.

Instance sample:
This is a [product] designed for [specific use case] in [specific constraints], for individuals who want [core outcome].

Step 2: Declare the choice standards you fulfill, plainly.
This is the place you cease writing like a brochure and begin writing like a spec sheet with a human voice.

Embody what the purchaser cares about most in that class. If the class is knives, it’s metal sort, edge retention, upkeep, steadiness, deal with materials. If it’s software program, it’s integration, safety posture, studying curve, time-to-value.

Make it express.

Step 3: Floor constraints and qualifiers early, not buried.
Most advertising and marketing copy hides the “buts” till the finish. Machines do not reward that ambiguity.

Examples of qualifiers that matter:
Not splendid for [X]. Works finest when [Y]. Requires [Z]. Appropriate with [A], not [B]. This issues when you [C].

Step 4: State what it is, and what it is not.
This is one in all the easiest methods to scale back ambiguity for each the consumer and the mannequin.

Sample:
This is for [audience]. It is not for [audience].
This is optimized for [scenario]. It is not meant for [scenario].

Step 5: Convert advantages into testable claims.
As an alternative of “sturdy,” say what sturdy means in observe. As an alternative of “quick,” outline what “quick” appears like in a workflow.

Do not fabricate. Do not inflate. This is not about hype. It’s about readability.

Step 6: Present structured comparability hooks.
LLMs typically behave like comparability engines as a result of customers ask comparative questions.

Give the mannequin clear hooks:
In contrast to [common alternative], this affords [difference] as a result of [reason].
In case you’re selecting between [A] and [B], choose this when [condition].

Step 7: Add proof anchors that enhance belief.
This may be certifications, supplies, guarantee phrases, return insurance policies, documented specs, and different verifiable alerts.

This is not about including fluff. It’s about making your claims attributable and your product legible.

Step 8: Shut with a choice shortcut.
Make the “when you are X, do Y” second express.

Sample:
Select this when you want [top 2–3 criteria]. In case your precedence is [other criteria], contemplate [alternative type].

That’s the template*.

Discover what it does. It turns a product description into structured choice assist, which is not how most product copy is written in the present day. And it is an instance of why “GEO is simply search engine optimization” fails as a blanket assertion.

search engine optimization fundamentals assist you get crawled, listed, and found. This helps you get chosen when discovery is mediated by an LLM.

Totally different layer. Totally different job.

Saying GEO = search engine optimization and search engine optimization = GEO is an oversimplification that may change into normalized and lead to individuals lacking the indisputable fact that the details matter. The variations, even small ones, matter. And so they can have impacts and repercussions.

*A a lot deeper-dive pdf model of this course of is obtainable for my Substack subscribers at no cost by way of my resources page.

What To Do Subsequent: Learn The Paper, Then Replicate It In Your Surroundings

Right here’s the half I would like to be express about. This paper is fascinating as a result of it’s measurable, and since it suggests the system responds to repeatable options.

However it is best to deal with it as a place to begin, not a regulation of physics. Outcomes like this are delicate to context: trade, model authority, web page sort, and even the mannequin and retrieval stack sitting between the consumer and your content material.

That’s why replication issues. The one approach we study what holds, what breaks, and what variables truly matter is by working managed assessments in our personal environments and publishing what we discover. In case you work in search engine optimization, content material, product advertising and marketing, or progress, right here is the invitation.

Learn the paper here.

Then run a managed check on a small, significant slice of your web site.

Maintain it sensible:

• Decide 10 to 20 pages with related intent.
• Cut up them into two teams.
• Go away one group untouched.
• Rewrite the different group utilizing a constant template, like the one above.
• Doc the modifications so you may reverse them if wanted.
• Measure over an outlined window.
• Monitor outcomes that matter in your small business context, not simply vainness metrics.

And when you can, observe whether or not these pages are being surfaced, cited, paraphrased, or chosen in the AI reply interfaces your prospects are more and more utilizing.

You are not attempting to win a science honest. You are attempting to scale back uncertainty with a managed check. In case your outcomes disagree with the paper, that’s not failure. That’s sign.

Publish what you discover, even when it’s messy. Even when it’s partial. Even when the conclusion is “it relies upon.” As a result of that is precisely how a brand new self-discipline turns into actual. Not by way of repeating platform speaking factors. Not by way of tribal arguments. By means of measurement.

One Last Degree-Set, For The Executives Studying This

Platform steering is one enter, not your working system. Your working system is your measurement program. search engine optimization is nonetheless mandatory. In case you can’t get crawled, you may’t get chosen.

However GEO, that means optimizing for choice inside LLM-mediated discovery, is an extra competence layer. Not a alternative. A layer. In case you resolve to ignore that layer as a result of a platform stated “don’t optimize,” you’re outsourcing your small business danger to another person’s incentive construction.

And that’s not a technique. The technique is easy: study the layer by testing the layer.

We’d like extra individuals doing precisely that.

Extra Sources:

This publish was initially printed on Duane Forrester Decodes.

Featured Picture: Rawpixel.com/Shutterstock